Machine Learning Models
Artificial intelligence (AI) is a buzz word that everyone has come across no matter which domain they work in - AI powers cars that drive themselves, public schools in the US use AI to identify dropouts and warn teachers to take action, French authorities use an AI system to identify pool owners who haven’t paid taxes, AI-powered inspection systems are becoming the norm in most manufacturing processes, insurance claims are increasingly assessed and processed with AI systems, in the agriculture sector AI is used to quantify crop yield and detect plant diseases, AI is used to recognize misinformation spread in social media, and the list goes on.
At the core of all these applications is Machine Learning – an AI tool that learns patterns from data and gathers insights. Let’s try to understand this with a human example:
Students of radiology are given a CT scan to identify infections of the lung. With the help of a trainer, the students will identify patterns in the scan that are characteristic of that infection and this will allow them to identify the same pattern in other patients. When the students come across a scan of a different kind of infection, they might fail to identify it because the scan doesn’t show all characteristics of the infection they were trained on. Again, with the help of the trainer, the students will learn about this new pattern as well. Over time, after having seen and learnt several patterns of lung CT scans, the students turn into experts who can apply their knowledge broadly and can identify various forms of infections in CT scans.
Machine learning is a process that closely resembles this human-like learning. One of the classical tools for machine learning is a neural network. A machine can only learn when a neural network is specifically designed for the task and trained. During the training process, the strength of the connection between two neurons – also known as ‘weights’ are updated to fit the data. Let us reconsider our example: to identify lung infections from CT scans, just like the students, the neural network is provided lots of images of lung CT scans along with the label of the infection. The network then goes through a training process adjusting the weights of the neurons until the correct patterns are learnt. When a new scan showing a different kind of infection is provided to the network, it will fail to identify the infection. The training process is then continued including this new set of images which adjusts the weights of the neurons once again. The architecture of the neural network along with the weights of the neurons constitute a machine learning model which captures all the insights from the training data. A machine learning model can be considered as a mathematical representation of the real-world data generating process. A well-trained ML model has learnt from a wide variety of data and can efficiently detect these patterns in new and unknown data without any human intervention.
At ARRALYZE, we are developing various machine learning models that will automate processes and increase throughput in our high throughput screening set up. We have developed models that investigate the microscopic wells present in the glass chips distinguishing live cells from dead cells and counting the number of cells in each well as well as tracking these cells over a period of time, helping to produce biological insights.